Hi @minjoy ,
This is a good question, but one that does not have a simple answer. There are many ways to test, e.g., using simulated datasets or real datasets with a known composition (e.g., from a mock community). I would recommend using a mock community — you can see some discussion and resources for using mock communities here:
and an old repository of mock community datasets that you can use for the purpose here:
This is not a short process: you will need to pick a mock community, process the data with QIIME 2, and then evaluate the accuracy of your classifier (see the thread linked above for relevant discussion and methods). But once you have a working mock community you can use it for testing any classifiers that you make.
Good luck!